A team of researchers gains access to secret OpenAI algorithms using Anthropic AI

A group of cybersecurity researchers has successfully gained access to internal systems at OpenAI using Anthropic’s AI model, Claude. The exploit was carried out through a corporate ChatGPT account belonging to an OpenAI employee, allowing the team to view internal files and propose changes to the company’s proprietary algorithm repository. This unauthorized access occurred during a sanctioned bug-bounty program run by OpenAI, a common practice in the tech industry designed to identify system weaknesses before malicious actors can exploit them. As a result, the researchers were compensated with $6,500 for their findings. The attack unfolded in early July, with the research team from Hacktron AI identifying a flaw in Discourse, the external platform hosting OpenAI’s community forum. Initially, they attempted to exploit the vulnerability using a specialized version of Claude Opus 4.8, but the attempt failed. However, following the release of Opus 5 by Anthropic later that same day, the team successfully exploited the weakness within three hours. With access to user credentials obtained through Discourse, they were able to log into ChatGPT accounts and penetrate OpenAI’s internal systems. Through this entry point, they reached Monorepo, the central repository where OpenAI stores some of its core algorithms, which are critical to the performance of its AI models. While the researchers did not manage to access the inner workings of ChatG’ts parameters, those determining how information is processed and responses generated, they were able to review official documentation and test the extent of their access. To demonstrate their intrusion, they submitted a request to modify the code under their team’s name. OpenAI acknowledged the breach, stating it had limited the permissions of community login tokens and revoked affected tokens and sessions. This incident adds to growing concerns over the security of large language models. Mohan Pedhapati, technology director at Hacktron AI, told The Wall Street Journal that while the team was not as powerful as Chinese threat actors, their ability to exploit the system highlights ongoing vulnerabilities. He noted that the reference to Chinese hackers alludes to a separate incident involving Anthropic, where a real attack from China reportedly breached its systems. The breach underscores the increasing sophistication of AI-driven security threats. Researchers from Hacktron AI emphasized that the entire process, from initial discovery to gaining access to OpenAI’s repositories, took less than 72 hours. Their disclosure came in a blog post, which was initially shared with The Wall Street Journal. OpenAI confirmed the breach and expressed gratitude for the researchers’ prompt reporting and cooperation in addressing the issue. Anthropic, meanwhile, has not issued a formal statement regarding the incident. However, the company previously raised alarms about the risks posed by AI, including the Hugging Face breach in July, where OpenAI’s AI systems allegedly collaborated to infiltrate another AI developer’s network. In a recent interview, Anthropic CEO Dario Amodei warned of “real dangers” surrounding AI and called for a slowdown in its development to mitigate potential harms. His comments align with broader calls within the tech industry for more cautious approaches to AI innovation.